Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-sse2-u24.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/sse2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/vcvt.h"
17
18
19 void xnn_qu8_f32_vcvt_ukernel__sse2_u24(
20 size_t batch,
21 const uint8_t* input,
22 float* output,
23 const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m128i vmagic_exp = _mm_set1_epi16(UINT16_C(0x4B00));
31 const __m128 vmagic_bias = _mm_set1_ps((float) (INT32_C(0x00800000) + (int32_t) params->scalar.zero_point));
32 const __m128 vscale = _mm_set1_ps(params->scalar.scale);
33 const __m128i vzero = _mm_setzero_si128();
34 XNN_FORCE_REALIZATION(vmagic_exp);
35 XNN_FORCE_REALIZATION(vmagic_bias);
36 XNN_FORCE_REALIZATION(vscale);
37 for (; batch >= 24 * sizeof(uint8_t); batch -= 24 * sizeof(uint8_t)) {
38 __m128i vx01234567 = _mm_loadl_epi64((const __m128i*) input);
39 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (input + 8));
40 __m128i vxGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input + 16));
41 input += 24;
42
43
44 vx01234567 = _mm_unpacklo_epi8(vx01234567, vzero);
45 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero);
46 vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, vzero);
47
48 __m128 vy0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(vx01234567, vmagic_exp));
49 __m128 vy4567 = _mm_castsi128_ps(_mm_unpackhi_epi16(vx01234567, vmagic_exp));
50 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp));
51 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp));
52 __m128 vyGHIJ = _mm_castsi128_ps(_mm_unpacklo_epi16(vxGHIJKLMN, vmagic_exp));
53 __m128 vyKLMN = _mm_castsi128_ps(_mm_unpackhi_epi16(vxGHIJKLMN, vmagic_exp));
54
55 vy0123 = _mm_sub_ps(vy0123, vmagic_bias);
56 vy4567 = _mm_sub_ps(vy4567, vmagic_bias);
57 vy89AB = _mm_sub_ps(vy89AB, vmagic_bias);
58 vyCDEF = _mm_sub_ps(vyCDEF, vmagic_bias);
59 vyGHIJ = _mm_sub_ps(vyGHIJ, vmagic_bias);
60 vyKLMN = _mm_sub_ps(vyKLMN, vmagic_bias);
61
62 vy0123 = _mm_mul_ps(vy0123, vscale);
63 vy4567 = _mm_mul_ps(vy4567, vscale);
64 vy89AB = _mm_mul_ps(vy89AB, vscale);
65 vyCDEF = _mm_mul_ps(vyCDEF, vscale);
66 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale);
67 vyKLMN = _mm_mul_ps(vyKLMN, vscale);
68
69 _mm_storeu_ps(output, vy0123);
70 _mm_storeu_ps(output + 4, vy4567);
71 _mm_storeu_ps(output + 8, vy89AB);
72 _mm_storeu_ps(output + 12, vyCDEF);
73 _mm_storeu_ps(output + 16, vyGHIJ);
74 _mm_storeu_ps(output + 20, vyKLMN);
75 output += 24;
76 }
77 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
78 __m128i vx = _mm_loadl_epi64((const __m128i*) input);
79 vx = _mm_unpacklo_epi8(vx, vzero);
80 input += 8;
81
82 __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
83 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
84
85 vy_lo = _mm_sub_ps(vy_lo, vmagic_bias);
86 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias);
87
88 vy_lo = _mm_mul_ps(vy_lo, vscale);
89 vy_hi = _mm_mul_ps(vy_hi, vscale);
90
91 _mm_storeu_ps(output, vy_lo);
92 _mm_storeu_ps(output + 4, vy_hi);
93 output += 8;
94 }
95 if XNN_UNLIKELY(batch != 0) {
96 assert(batch >= 1 * sizeof(uint8_t));
97 assert(batch <= 7 * sizeof(uint8_t));
98
99 __m128i vx = _mm_loadl_epi64((const __m128i*) input);
100 vx = _mm_unpacklo_epi8(vx, vzero);
101
102 __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
103 vy = _mm_sub_ps(vy, vmagic_bias);
104 vy = _mm_mul_ps(vy, vscale);
105
106 if (batch & (4 * sizeof(uint8_t))) {
107 _mm_storeu_ps(output, vy);
108 vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
109 vy = _mm_sub_ps(vy, vmagic_bias);
110 vy = _mm_mul_ps(vy, vscale);
111 output += 4;
112 }
113 if (batch & (2 * sizeof(uint8_t))) {
114 _mm_storel_pi((__m64*) output, vy);
115 vy = _mm_movehl_ps(vy, vy);
116 output += 2;
117 }
118 if (batch & (1 * sizeof(uint8_t))) {
119 _mm_store_ss(output, vy);
120 }
121 }
122 }